pszemraj commited on
Commit
bfb6abe
1 Parent(s): e24733e

load model from drive and convert

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
README.md ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ tags:
4
+ - generated_from_trainer
5
+ model-index:
6
+ - name: opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8
7
+ results: []
8
+ ---
9
+
10
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
11
+ should probably proofread and complete it, then remove this comment. -->
12
+
13
+ # opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8
14
+
15
+ This model is a fine-tuned version of [facebook/opt-2.7b](https://huggingface.co/facebook/opt-2.7b) on the None dataset.
16
+
17
+ ## Model description
18
+
19
+ More information needed
20
+
21
+ ## Intended uses & limitations
22
+
23
+ More information needed
24
+
25
+ ## Training and evaluation data
26
+
27
+ More information needed
28
+
29
+ ## Training procedure
30
+
31
+ ### Training hyperparameters
32
+
33
+ The following hyperparameters were used during training:
34
+ - learning_rate: 4e-05
35
+ - train_batch_size: 8
36
+ - eval_batch_size: 8
37
+ - seed: 42
38
+ - distributed_type: multi-GPU
39
+ - gradient_accumulation_steps: 16
40
+ - total_train_batch_size: 128
41
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
42
+ - lr_scheduler_type: cosine
43
+ - lr_scheduler_warmup_ratio: 0.01
44
+ - num_epochs: 3
45
+
46
+ ### Training results
47
+
48
+
49
+
50
+ ### Framework versions
51
+
52
+ - Transformers 4.19.2
53
+ - Pytorch 1.10.0+cu113
54
+ - Datasets 2.2.2
55
+ - Tokenizers 0.12.1
config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/opt-2.7b",
3
+ "activation_dropout": 0.0,
4
+ "activation_function": "relu",
5
+ "architectures": [
6
+ "OPTForCausalLM"
7
+ ],
8
+ "attention_dropout": 0.0,
9
+ "bos_token_id": 0,
10
+ "do_layer_norm_before": true,
11
+ "dropout": 0.1,
12
+ "eos_token_id": 2,
13
+ "ffn_dim": 10240,
14
+ "hidden_size": 2560,
15
+ "init_std": 0.02,
16
+ "layerdrop": 0.0,
17
+ "max_position_embeddings": 2048,
18
+ "model_type": "opt",
19
+ "num_attention_heads": 32,
20
+ "num_hidden_layers": 32,
21
+ "pad_token_id": 1,
22
+ "prefix": "</s>",
23
+ "torch_dtype": "bfloat16",
24
+ "transformers_version": "4.19.2",
25
+ "use_cache": false,
26
+ "vocab_size": 50265,
27
+ "word_embed_proj_dim": 2560
28
+ }
latest ADDED
@@ -0,0 +1 @@
 
 
1
+ global_step1852
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8_training_metadata.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"output_dir": "/content/drive/MyDrive/Programming/hf-trainer/opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "overwrite_output_dir": true, "do_train": false, "do_eval": false, "do_predict": false, "evaluation_strategy": "no", "prediction_loss_only": false, "per_device_train_batch_size": 8, "per_device_eval_batch_size": 8, "per_gpu_train_batch_size": "None", "per_gpu_eval_batch_size": "None", "gradient_accumulation_steps": 16, "eval_accumulation_steps": 8, "eval_delay": 0, "learning_rate": 4e-05, "weight_decay": 0.1, "adam_beta1": 0.9, "adam_beta2": 0.999, "adam_epsilon": 1e-08, "max_grad_norm": 1, "num_train_epochs": 3, "max_steps": -1, "lr_scheduler_type": "cosine", "warmup_ratio": 0.01, "warmup_steps": 0, "log_level": -1, "log_level_replica": -1, "log_on_each_node": true, "logging_dir": "/content/drive/MyDrive/Programming/hf-trainer/opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8/logs", "logging_strategy": "steps", "logging_first_step": false, "logging_steps": 5, "logging_nan_inf_filter": true, "save_strategy": "epoch", "save_steps": 500, "save_total_limit": 1, "save_on_each_node": false, "no_cuda": false, "seed": 42, "data_seed": "None", "bf16": true, "fp16": false, "fp16_opt_level": "O1", "half_precision_backend": "amp", "bf16_full_eval": true, "fp16_full_eval": false, "tf32": "None", "local_rank": 0, "xpu_backend": "None", "tpu_num_cores": "None", "tpu_metrics_debug": false, "debug": "[]", "dataloader_drop_last": false, "eval_steps": "None", "dataloader_num_workers": 0, "past_index": -1, "run_name": "/content/drive/MyDrive/Programming/hf-trainer/opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "disable_tqdm": false, "remove_unused_columns": true, "label_names": "None", "load_best_model_at_end": false, "metric_for_best_model": "None", "greater_is_better": "None", "ignore_data_skip": false, "sharded_ddp": "[]", "fsdp": "[]", "fsdp_min_num_params": 0, "deepspeed": "ds_config_zero2_bf16.json", "label_smoothing_factor": 0.0, "optim": "adamw_hf", "adafactor": false, "group_by_length": false, "length_column_name": "length", "report_to": "['tensorboard']", "ddp_find_unused_parameters": "None", "ddp_bucket_cap_mb": "None", "dataloader_pin_memory": true, "skip_memory_metrics": true, "use_legacy_prediction_loop": false, "push_to_hub": true, "resume_from_checkpoint": "None", "hub_model_id": "opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "hub_strategy": "end", "hub_token": "<HUB_TOKEN>", "hub_private_repo": false, "gradient_checkpointing": true, "include_inputs_for_metrics": false, "fp16_backend": "auto", "push_to_hub_model_id": "None", "push_to_hub_organization": "None", "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>", "_n_gpu": 1, "mp_parameters": "", "auto_find_batch_size": false, "full_determinism": false, "train_batch_size": 8, "eval_batch_size": 8, "configs_src": "opt-2pt7b-ps_DS-msgs_Ep-3_Bs-8", "data_tag": "text-file-input"}
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:69aff7264edf04861c8f3b6eed3f553e343af5ee8aebeb5ad1635f39ad2b4683
3
+ size 10606359699
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "eos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "unk_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "pad_token": "</s>"}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "bos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "eos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_prefix_space": false, "errors": "replace", "pad_token": {"content": "<pad>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_bos_token": true, "special_tokens_map_file": null, "name_or_path": "facebook/opt-2.7b", "model_max_length": 512, "tokenizer_class": "GPT2Tokenizer"}
trainer_state.json ADDED
@@ -0,0 +1,2245 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": null,
3
+ "best_model_checkpoint": null,
4
+ "epoch": 2.9985838559579205,
5
+ "global_step": 1851,
6
+ "is_hyper_param_search": false,
7
+ "is_local_process_zero": true,
8
+ "is_world_process_zero": true,
9
+ "log_history": [
10
+ {
11
+ "epoch": 0.01,
12
+ "learning_rate": 1.0526315789473684e-05,
13
+ "loss": 11.5,
14
+ "step": 5
15
+ },
16
+ {
17
+ "epoch": 0.02,
18
+ "learning_rate": 2.105263157894737e-05,
19
+ "loss": 6.7393,
20
+ "step": 10
21
+ },
22
+ {
23
+ "epoch": 0.02,
24
+ "learning_rate": 3.157894736842106e-05,
25
+ "loss": 3.7191,
26
+ "step": 15
27
+ },
28
+ {
29
+ "epoch": 0.03,
30
+ "learning_rate": 3.999997059313686e-05,
31
+ "loss": 3.1877,
32
+ "step": 20
33
+ },
34
+ {
35
+ "epoch": 0.04,
36
+ "learning_rate": 3.999894136200706e-05,
37
+ "loss": 2.9055,
38
+ "step": 25
39
+ },
40
+ {
41
+ "epoch": 0.05,
42
+ "learning_rate": 3.9996441874195635e-05,
43
+ "loss": 2.9141,
44
+ "step": 30
45
+ },
46
+ {
47
+ "epoch": 0.06,
48
+ "learning_rate": 3.999247231345674e-05,
49
+ "loss": 2.7834,
50
+ "step": 35
51
+ },
52
+ {
53
+ "epoch": 0.06,
54
+ "learning_rate": 3.998703297161948e-05,
55
+ "loss": 2.7901,
56
+ "step": 40
57
+ },
58
+ {
59
+ "epoch": 0.07,
60
+ "learning_rate": 3.9980124248566466e-05,
61
+ "loss": 2.8676,
62
+ "step": 45
63
+ },
64
+ {
65
+ "epoch": 0.08,
66
+ "learning_rate": 3.9971746652204386e-05,
67
+ "loss": 2.6787,
68
+ "step": 50
69
+ },
70
+ {
71
+ "epoch": 0.09,
72
+ "learning_rate": 3.996190079842669e-05,
73
+ "loss": 2.704,
74
+ "step": 55
75
+ },
76
+ {
77
+ "epoch": 0.1,
78
+ "learning_rate": 3.995058741106831e-05,
79
+ "loss": 2.6286,
80
+ "step": 60
81
+ },
82
+ {
83
+ "epoch": 0.11,
84
+ "learning_rate": 3.993780732185244e-05,
85
+ "loss": 2.6983,
86
+ "step": 65
87
+ },
88
+ {
89
+ "epoch": 0.11,
90
+ "learning_rate": 3.992356147032939e-05,
91
+ "loss": 2.6522,
92
+ "step": 70
93
+ },
94
+ {
95
+ "epoch": 0.12,
96
+ "learning_rate": 3.9907850903807514e-05,
97
+ "loss": 2.6293,
98
+ "step": 75
99
+ },
100
+ {
101
+ "epoch": 0.13,
102
+ "learning_rate": 3.989067677727622e-05,
103
+ "loss": 2.5399,
104
+ "step": 80
105
+ },
106
+ {
107
+ "epoch": 0.14,
108
+ "learning_rate": 3.987204035332105e-05,
109
+ "loss": 2.6687,
110
+ "step": 85
111
+ },
112
+ {
113
+ "epoch": 0.15,
114
+ "learning_rate": 3.985194300203087e-05,
115
+ "loss": 2.6291,
116
+ "step": 90
117
+ },
118
+ {
119
+ "epoch": 0.15,
120
+ "learning_rate": 3.983038620089714e-05,
121
+ "loss": 2.6533,
122
+ "step": 95
123
+ },
124
+ {
125
+ "epoch": 0.16,
126
+ "learning_rate": 3.980737153470528e-05,
127
+ "loss": 2.5923,
128
+ "step": 100
129
+ },
130
+ {
131
+ "epoch": 0.17,
132
+ "learning_rate": 3.97829006954182e-05,
133
+ "loss": 2.5941,
134
+ "step": 105
135
+ },
136
+ {
137
+ "epoch": 0.18,
138
+ "learning_rate": 3.9756975482051855e-05,
139
+ "loss": 2.6612,
140
+ "step": 110
141
+ },
142
+ {
143
+ "epoch": 0.19,
144
+ "learning_rate": 3.972959780054306e-05,
145
+ "loss": 2.6058,
146
+ "step": 115
147
+ },
148
+ {
149
+ "epoch": 0.19,
150
+ "learning_rate": 3.9700769663609304e-05,
151
+ "loss": 2.5226,
152
+ "step": 120
153
+ },
154
+ {
155
+ "epoch": 0.2,
156
+ "learning_rate": 3.967049319060081e-05,
157
+ "loss": 2.5573,
158
+ "step": 125
159
+ },
160
+ {
161
+ "epoch": 0.21,
162
+ "learning_rate": 3.963877060734473e-05,
163
+ "loss": 2.603,
164
+ "step": 130
165
+ },
166
+ {
167
+ "epoch": 0.22,
168
+ "learning_rate": 3.9605604245981515e-05,
169
+ "loss": 2.6506,
170
+ "step": 135
171
+ },
172
+ {
173
+ "epoch": 0.23,
174
+ "learning_rate": 3.9570996544793445e-05,
175
+ "loss": 2.631,
176
+ "step": 140
177
+ },
178
+ {
179
+ "epoch": 0.23,
180
+ "learning_rate": 3.9534950048025396e-05,
181
+ "loss": 2.5668,
182
+ "step": 145
183
+ },
184
+ {
185
+ "epoch": 0.24,
186
+ "learning_rate": 3.9497467405697756e-05,
187
+ "loss": 2.6354,
188
+ "step": 150
189
+ },
190
+ {
191
+ "epoch": 0.25,
192
+ "learning_rate": 3.9458551373411664e-05,
193
+ "loss": 2.5427,
194
+ "step": 155
195
+ },
196
+ {
197
+ "epoch": 0.26,
198
+ "learning_rate": 3.941820481214637e-05,
199
+ "loss": 2.5586,
200
+ "step": 160
201
+ },
202
+ {
203
+ "epoch": 0.27,
204
+ "learning_rate": 3.937643068804896e-05,
205
+ "loss": 2.5577,
206
+ "step": 165
207
+ },
208
+ {
209
+ "epoch": 0.28,
210
+ "learning_rate": 3.933323207221624e-05,
211
+ "loss": 2.5664,
212
+ "step": 170
213
+ },
214
+ {
215
+ "epoch": 0.28,
216
+ "learning_rate": 3.9288612140468984e-05,
217
+ "loss": 2.5396,
218
+ "step": 175
219
+ },
220
+ {
221
+ "epoch": 0.29,
222
+ "learning_rate": 3.924257417311846e-05,
223
+ "loss": 2.5558,
224
+ "step": 180
225
+ },
226
+ {
227
+ "epoch": 0.3,
228
+ "learning_rate": 3.919512155472529e-05,
229
+ "loss": 2.5306,
230
+ "step": 185
231
+ },
232
+ {
233
+ "epoch": 0.31,
234
+ "learning_rate": 3.9146257773850585e-05,
235
+ "loss": 2.4945,
236
+ "step": 190
237
+ },
238
+ {
239
+ "epoch": 0.32,
240
+ "learning_rate": 3.9095986422799506e-05,
241
+ "loss": 2.6086,
242
+ "step": 195
243
+ },
244
+ {
245
+ "epoch": 0.32,
246
+ "learning_rate": 3.904431119735718e-05,
247
+ "loss": 2.4973,
248
+ "step": 200
249
+ },
250
+ {
251
+ "epoch": 0.33,
252
+ "learning_rate": 3.899123589651695e-05,
253
+ "loss": 2.5872,
254
+ "step": 205
255
+ },
256
+ {
257
+ "epoch": 0.34,
258
+ "learning_rate": 3.893676442220114e-05,
259
+ "loss": 2.5216,
260
+ "step": 210
261
+ },
262
+ {
263
+ "epoch": 0.35,
264
+ "learning_rate": 3.888090077897418e-05,
265
+ "loss": 2.5367,
266
+ "step": 215
267
+ },
268
+ {
269
+ "epoch": 0.36,
270
+ "learning_rate": 3.882364907374819e-05,
271
+ "loss": 2.5495,
272
+ "step": 220
273
+ },
274
+ {
275
+ "epoch": 0.36,
276
+ "learning_rate": 3.8765013515481065e-05,
277
+ "loss": 2.6037,
278
+ "step": 225
279
+ },
280
+ {
281
+ "epoch": 0.37,
282
+ "learning_rate": 3.870499841486705e-05,
283
+ "loss": 2.5012,
284
+ "step": 230
285
+ },
286
+ {
287
+ "epoch": 0.38,
288
+ "learning_rate": 3.864360818401982e-05,
289
+ "loss": 2.4773,
290
+ "step": 235
291
+ },
292
+ {
293
+ "epoch": 0.39,
294
+ "learning_rate": 3.8580847336148105e-05,
295
+ "loss": 2.52,
296
+ "step": 240
297
+ },
298
+ {
299
+ "epoch": 0.4,
300
+ "learning_rate": 3.851672048522395e-05,
301
+ "loss": 2.4718,
302
+ "step": 245
303
+ },
304
+ {
305
+ "epoch": 0.4,
306
+ "learning_rate": 3.845123234564343e-05,
307
+ "loss": 2.5251,
308
+ "step": 250
309
+ },
310
+ {
311
+ "epoch": 0.41,
312
+ "learning_rate": 3.838438773188014e-05,
313
+ "loss": 2.5547,
314
+ "step": 255
315
+ },
316
+ {
317
+ "epoch": 0.42,
318
+ "learning_rate": 3.831619155813119e-05,
319
+ "loss": 2.475,
320
+ "step": 260
321
+ },
322
+ {
323
+ "epoch": 0.43,
324
+ "learning_rate": 3.8246648837955965e-05,
325
+ "loss": 2.4957,
326
+ "step": 265
327
+ },
328
+ {
329
+ "epoch": 0.44,
330
+ "learning_rate": 3.817576468390753e-05,
331
+ "loss": 2.5425,
332
+ "step": 270
333
+ },
334
+ {
335
+ "epoch": 0.45,
336
+ "learning_rate": 3.810354430715678e-05,
337
+ "loss": 2.5495,
338
+ "step": 275
339
+ },
340
+ {
341
+ "epoch": 0.45,
342
+ "learning_rate": 3.802999301710932e-05,
343
+ "loss": 2.5375,
344
+ "step": 280
345
+ },
346
+ {
347
+ "epoch": 0.46,
348
+ "learning_rate": 3.795511622101516e-05,
349
+ "loss": 2.5151,
350
+ "step": 285
351
+ },
352
+ {
353
+ "epoch": 0.47,
354
+ "learning_rate": 3.787891942357115e-05,
355
+ "loss": 2.4864,
356
+ "step": 290
357
+ },
358
+ {
359
+ "epoch": 0.48,
360
+ "learning_rate": 3.780140822651633e-05,
361
+ "loss": 2.5396,
362
+ "step": 295
363
+ },
364
+ {
365
+ "epoch": 0.49,
366
+ "learning_rate": 3.77225883282201e-05,
367
+ "loss": 2.5585,
368
+ "step": 300
369
+ },
370
+ {
371
+ "epoch": 0.49,
372
+ "learning_rate": 3.764246552326328e-05,
373
+ "loss": 2.4947,
374
+ "step": 305
375
+ },
376
+ {
377
+ "epoch": 0.5,
378
+ "learning_rate": 3.756104570201213e-05,
379
+ "loss": 2.5004,
380
+ "step": 310
381
+ },
382
+ {
383
+ "epoch": 0.51,
384
+ "learning_rate": 3.747833485018529e-05,
385
+ "loss": 2.4729,
386
+ "step": 315
387
+ },
388
+ {
389
+ "epoch": 0.52,
390
+ "learning_rate": 3.739433904841375e-05,
391
+ "loss": 2.4764,
392
+ "step": 320
393
+ },
394
+ {
395
+ "epoch": 0.53,
396
+ "learning_rate": 3.7309064471793794e-05,
397
+ "loss": 2.5369,
398
+ "step": 325
399
+ },
400
+ {
401
+ "epoch": 0.53,
402
+ "learning_rate": 3.7222517389433085e-05,
403
+ "loss": 2.4836,
404
+ "step": 330
405
+ },
406
+ {
407
+ "epoch": 0.54,
408
+ "learning_rate": 3.7134704163989705e-05,
409
+ "loss": 2.4558,
410
+ "step": 335
411
+ },
412
+ {
413
+ "epoch": 0.55,
414
+ "learning_rate": 3.7045631251204434e-05,
415
+ "loss": 2.5562,
416
+ "step": 340
417
+ },
418
+ {
419
+ "epoch": 0.56,
420
+ "learning_rate": 3.6955305199426164e-05,
421
+ "loss": 2.4603,
422
+ "step": 345
423
+ },
424
+ {
425
+ "epoch": 0.57,
426
+ "learning_rate": 3.6863732649130426e-05,
427
+ "loss": 2.4176,
428
+ "step": 350
429
+ },
430
+ {
431
+ "epoch": 0.57,
432
+ "learning_rate": 3.677092033243128e-05,
433
+ "loss": 2.4991,
434
+ "step": 355
435
+ },
436
+ {
437
+ "epoch": 0.58,
438
+ "learning_rate": 3.667687507258631e-05,
439
+ "loss": 2.4846,
440
+ "step": 360
441
+ },
442
+ {
443
+ "epoch": 0.59,
444
+ "learning_rate": 3.658160378349508e-05,
445
+ "loss": 2.5998,
446
+ "step": 365
447
+ },
448
+ {
449
+ "epoch": 0.6,
450
+ "learning_rate": 3.648511346919079e-05,
451
+ "loss": 2.5061,
452
+ "step": 370
453
+ },
454
+ {
455
+ "epoch": 0.61,
456
+ "learning_rate": 3.638741122332539e-05,
457
+ "loss": 2.4538,
458
+ "step": 375
459
+ },
460
+ {
461
+ "epoch": 0.62,
462
+ "learning_rate": 3.628850422864807e-05,
463
+ "loss": 2.498,
464
+ "step": 380
465
+ },
466
+ {
467
+ "epoch": 0.62,
468
+ "learning_rate": 3.618839975647718e-05,
469
+ "loss": 2.4867,
470
+ "step": 385
471
+ },
472
+ {
473
+ "epoch": 0.63,
474
+ "learning_rate": 3.608710516616575e-05,
475
+ "loss": 2.5426,
476
+ "step": 390
477
+ },
478
+ {
479
+ "epoch": 0.64,
480
+ "learning_rate": 3.598462790456035e-05,
481
+ "loss": 2.4842,
482
+ "step": 395
483
+ },
484
+ {
485
+ "epoch": 0.65,
486
+ "learning_rate": 3.588097550545368e-05,
487
+ "loss": 2.4274,
488
+ "step": 400
489
+ },
490
+ {
491
+ "epoch": 0.66,
492
+ "learning_rate": 3.5776155589030725e-05,
493
+ "loss": 2.5354,
494
+ "step": 405
495
+ },
496
+ {
497
+ "epoch": 0.66,
498
+ "learning_rate": 3.5670175861308496e-05,
499
+ "loss": 2.455,
500
+ "step": 410
501
+ },
502
+ {
503
+ "epoch": 0.67,
504
+ "learning_rate": 3.556304411356954e-05,
505
+ "loss": 2.5039,
506
+ "step": 415
507
+ },
508
+ {
509
+ "epoch": 0.68,
510
+ "learning_rate": 3.545476822178915e-05,
511
+ "loss": 2.518,
512
+ "step": 420
513
+ },
514
+ {
515
+ "epoch": 0.69,
516
+ "learning_rate": 3.5345356146056326e-05,
517
+ "loss": 2.4718,
518
+ "step": 425
519
+ },
520
+ {
521
+ "epoch": 0.7,
522
+ "learning_rate": 3.52348159299886e-05,
523
+ "loss": 2.4318,
524
+ "step": 430
525
+ },
526
+ {
527
+ "epoch": 0.7,
528
+ "learning_rate": 3.512315570014071e-05,
529
+ "loss": 2.5146,
530
+ "step": 435
531
+ },
532
+ {
533
+ "epoch": 0.71,
534
+ "learning_rate": 3.50103836654071e-05,
535
+ "loss": 2.4801,
536
+ "step": 440
537
+ },
538
+ {
539
+ "epoch": 0.72,
540
+ "learning_rate": 3.489650811641849e-05,
541
+ "loss": 2.4429,
542
+ "step": 445
543
+ },
544
+ {
545
+ "epoch": 0.73,
546
+ "learning_rate": 3.478153742493235e-05,
547
+ "loss": 2.4654,
548
+ "step": 450
549
+ },
550
+ {
551
+ "epoch": 0.74,
552
+ "learning_rate": 3.4665480043217444e-05,
553
+ "loss": 2.4846,
554
+ "step": 455
555
+ },
556
+ {
557
+ "epoch": 0.74,
558
+ "learning_rate": 3.454834450343245e-05,
559
+ "loss": 2.4371,
560
+ "step": 460
561
+ },
562
+ {
563
+ "epoch": 0.75,
564
+ "learning_rate": 3.443013941699868e-05,
565
+ "loss": 2.4071,
566
+ "step": 465
567
+ },
568
+ {
569
+ "epoch": 0.76,
570
+ "learning_rate": 3.431087347396702e-05,
571
+ "loss": 2.3886,
572
+ "step": 470
573
+ },
574
+ {
575
+ "epoch": 0.77,
576
+ "learning_rate": 3.419055544237906e-05,
577
+ "loss": 2.4539,
578
+ "step": 475
579
+ },
580
+ {
581
+ "epoch": 0.78,
582
+ "learning_rate": 3.40691941676225e-05,
583
+ "loss": 2.4406,
584
+ "step": 480
585
+ },
586
+ {
587
+ "epoch": 0.78,
588
+ "learning_rate": 3.394679857178086e-05,
589
+ "loss": 2.3687,
590
+ "step": 485
591
+ },
592
+ {
593
+ "epoch": 0.79,
594
+ "learning_rate": 3.382337765297756e-05,
595
+ "loss": 2.4244,
596
+ "step": 490
597
+ },
598
+ {
599
+ "epoch": 0.8,
600
+ "learning_rate": 3.3698940484714394e-05,
601
+ "loss": 2.4477,
602
+ "step": 495
603
+ },
604
+ {
605
+ "epoch": 0.81,
606
+ "learning_rate": 3.35734962152045e-05,
607
+ "loss": 2.5432,
608
+ "step": 500
609
+ },
610
+ {
611
+ "epoch": 0.82,
612
+ "learning_rate": 3.34470540666998e-05,
613
+ "loss": 2.4633,
614
+ "step": 505
615
+ },
616
+ {
617
+ "epoch": 0.83,
618
+ "learning_rate": 3.331962333481302e-05,
619
+ "loss": 2.4302,
620
+ "step": 510
621
+ },
622
+ {
623
+ "epoch": 0.83,
624
+ "learning_rate": 3.319121338783428e-05,
625
+ "loss": 2.4528,
626
+ "step": 515
627
+ },
628
+ {
629
+ "epoch": 0.84,
630
+ "learning_rate": 3.3061833666042416e-05,
631
+ "loss": 2.3741,
632
+ "step": 520
633
+ },
634
+ {
635
+ "epoch": 0.85,
636
+ "learning_rate": 3.29314936810109e-05,
637
+ "loss": 2.4638,
638
+ "step": 525
639
+ },
640
+ {
641
+ "epoch": 0.86,
642
+ "learning_rate": 3.280020301490863e-05,
643
+ "loss": 2.4204,
644
+ "step": 530
645
+ },
646
+ {
647
+ "epoch": 0.87,
648
+ "learning_rate": 3.2667971319795473e-05,
649
+ "loss": 2.436,
650
+ "step": 535
651
+ },
652
+ {
653
+ "epoch": 0.87,
654
+ "learning_rate": 3.253480831691264e-05,
655
+ "loss": 2.4194,
656
+ "step": 540
657
+ },
658
+ {
659
+ "epoch": 0.88,
660
+ "learning_rate": 3.240072379596806e-05,
661
+ "loss": 2.3565,
662
+ "step": 545
663
+ },
664
+ {
665
+ "epoch": 0.89,
666
+ "learning_rate": 3.226572761441666e-05,
667
+ "loss": 2.4421,
668
+ "step": 550
669
+ },
670
+ {
671
+ "epoch": 0.9,
672
+ "learning_rate": 3.2129829696735636e-05,
673
+ "loss": 2.4169,
674
+ "step": 555
675
+ },
676
+ {
677
+ "epoch": 0.91,
678
+ "learning_rate": 3.1993040033694916e-05,
679
+ "loss": 2.4425,
680
+ "step": 560
681
+ },
682
+ {
683
+ "epoch": 0.91,
684
+ "learning_rate": 3.1855368681622584e-05,
685
+ "loss": 2.399,
686
+ "step": 565
687
+ },
688
+ {
689
+ "epoch": 0.92,
690
+ "learning_rate": 3.171682576166565e-05,
691
+ "loss": 2.3747,
692
+ "step": 570
693
+ },
694
+ {
695
+ "epoch": 0.93,
696
+ "learning_rate": 3.1577421459045905e-05,
697
+ "loss": 2.437,
698
+ "step": 575
699
+ },
700
+ {
701
+ "epoch": 0.94,
702
+ "learning_rate": 3.143716602231122e-05,
703
+ "loss": 2.4131,
704
+ "step": 580
705
+ },
706
+ {
707
+ "epoch": 0.95,
708
+ "learning_rate": 3.129606976258201e-05,
709
+ "loss": 2.4329,
710
+ "step": 585
711
+ },
712
+ {
713
+ "epoch": 0.95,
714
+ "learning_rate": 3.115414305279327e-05,
715
+ "loss": 2.4521,
716
+ "step": 590
717
+ },
718
+ {
719
+ "epoch": 0.96,
720
+ "learning_rate": 3.101139632693197e-05,
721
+ "loss": 2.3317,
722
+ "step": 595
723
+ },
724
+ {
725
+ "epoch": 0.97,
726
+ "learning_rate": 3.086784007926996e-05,
727
+ "loss": 2.4119,
728
+ "step": 600
729
+ },
730
+ {
731
+ "epoch": 0.98,
732
+ "learning_rate": 3.072348486359247e-05,
733
+ "loss": 2.4315,
734
+ "step": 605
735
+ },
736
+ {
737
+ "epoch": 0.99,
738
+ "learning_rate": 3.0578341292422286e-05,
739
+ "loss": 2.4687,
740
+ "step": 610
741
+ },
742
+ {
743
+ "epoch": 1.0,
744
+ "learning_rate": 3.043242003623947e-05,
745
+ "loss": 2.4703,
746
+ "step": 615
747
+ },
748
+ {
749
+ "epoch": 1.0,
750
+ "learning_rate": 3.0285731822696954e-05,
751
+ "loss": 2.5997,
752
+ "step": 620
753
+ },
754
+ {
755
+ "epoch": 1.01,
756
+ "learning_rate": 3.0138287435831855e-05,
757
+ "loss": 2.0103,
758
+ "step": 625
759
+ },
760
+ {
761
+ "epoch": 1.02,
762
+ "learning_rate": 2.9990097715272694e-05,
763
+ "loss": 1.9907,
764
+ "step": 630
765
+ },
766
+ {
767
+ "epoch": 1.03,
768
+ "learning_rate": 2.9841173555442463e-05,
769
+ "loss": 1.9292,
770
+ "step": 635
771
+ },
772
+ {
773
+ "epoch": 1.04,
774
+ "learning_rate": 2.9691525904757745e-05,
775
+ "loss": 1.9898,
776
+ "step": 640
777
+ },
778
+ {
779
+ "epoch": 1.05,
780
+ "learning_rate": 2.954116576482378e-05,
781
+ "loss": 1.9234,
782
+ "step": 645
783
+ },
784
+ {
785
+ "epoch": 1.05,
786
+ "learning_rate": 2.9390104189625702e-05,
787
+ "loss": 1.8726,
788
+ "step": 650
789
+ },
790
+ {
791
+ "epoch": 1.06,
792
+ "learning_rate": 2.923835228471587e-05,
793
+ "loss": 1.9208,
794
+ "step": 655
795
+ },
796
+ {
797
+ "epoch": 1.07,
798
+ "learning_rate": 2.90859212063974e-05,
799
+ "loss": 1.9407,
800
+ "step": 660
801
+ },
802
+ {
803
+ "epoch": 1.08,
804
+ "learning_rate": 2.8932822160904038e-05,
805
+ "loss": 1.9377,
806
+ "step": 665
807
+ },
808
+ {
809
+ "epoch": 1.09,
810
+ "learning_rate": 2.877906640357628e-05,
811
+ "loss": 1.9665,
812
+ "step": 670
813
+ },
814
+ {
815
+ "epoch": 1.09,
816
+ "learning_rate": 2.862466523803393e-05,
817
+ "loss": 1.9723,
818
+ "step": 675
819
+ },
820
+ {
821
+ "epoch": 1.1,
822
+ "learning_rate": 2.846963001534507e-05,
823
+ "loss": 1.9876,
824
+ "step": 680
825
+ },
826
+ {
827
+ "epoch": 1.11,
828
+ "learning_rate": 2.8313972133191615e-05,
829
+ "loss": 1.9405,
830
+ "step": 685
831
+ },
832
+ {
833
+ "epoch": 1.12,
834
+ "learning_rate": 2.8157703035031353e-05,
835
+ "loss": 1.9848,
836
+ "step": 690
837
+ },
838
+ {
839
+ "epoch": 1.13,
840
+ "learning_rate": 2.8000834209256665e-05,
841
+ "loss": 1.9328,
842
+ "step": 695
843
+ },
844
+ {
845
+ "epoch": 1.13,
846
+ "learning_rate": 2.7843377188349962e-05,
847
+ "loss": 1.9343,
848
+ "step": 700
849
+ },
850
+ {
851
+ "epoch": 1.14,
852
+ "learning_rate": 2.768534354803581e-05,
853
+ "loss": 1.9245,
854
+ "step": 705
855
+ },
856
+ {
857
+ "epoch": 1.15,
858
+ "learning_rate": 2.752674490642996e-05,
859
+ "loss": 1.9526,
860
+ "step": 710
861
+ },
862
+ {
863
+ "epoch": 1.16,
864
+ "learning_rate": 2.7367592923185207e-05,
865
+ "loss": 1.93,
866
+ "step": 715
867
+ },
868
+ {
869
+ "epoch": 1.17,
870
+ "learning_rate": 2.720789929863421e-05,
871
+ "loss": 1.9263,
872
+ "step": 720
873
+ },
874
+ {
875
+ "epoch": 1.17,
876
+ "learning_rate": 2.7047675772929328e-05,
877
+ "loss": 1.9432,
878
+ "step": 725
879
+ },
880
+ {
881
+ "epoch": 1.18,
882
+ "learning_rate": 2.6886934125179504e-05,
883
+ "loss": 1.9481,
884
+ "step": 730
885
+ },
886
+ {
887
+ "epoch": 1.19,
888
+ "learning_rate": 2.672568617258432e-05,
889
+ "loss": 1.909,
890
+ "step": 735
891
+ },
892
+ {
893
+ "epoch": 1.2,
894
+ "learning_rate": 2.6563943769565258e-05,
895
+ "loss": 1.9386,
896
+ "step": 740
897
+ },
898
+ {
899
+ "epoch": 1.21,
900
+ "learning_rate": 2.6401718806894144e-05,
901
+ "loss": 1.9362,
902
+ "step": 745
903
+ },
904
+ {
905
+ "epoch": 1.22,
906
+ "learning_rate": 2.6239023210819027e-05,
907
+ "loss": 1.9494,
908
+ "step": 750
909
+ },
910
+ {
911
+ "epoch": 1.22,
912
+ "learning_rate": 2.6075868942187366e-05,
913
+ "loss": 1.9576,
914
+ "step": 755
915
+ },
916
+ {
917
+ "epoch": 1.23,
918
+ "learning_rate": 2.5912267995566746e-05,
919
+ "loss": 1.937,
920
+ "step": 760
921
+ },
922
+ {
923
+ "epoch": 1.24,
924
+ "learning_rate": 2.5748232398363044e-05,
925
+ "loss": 1.9889,
926
+ "step": 765
927
+ },
928
+ {
929
+ "epoch": 1.25,
930
+ "learning_rate": 2.5583774209936218e-05,
931
+ "loss": 1.9285,
932
+ "step": 770
933
+ },
934
+ {
935
+ "epoch": 1.26,
936
+ "learning_rate": 2.5418905520713767e-05,
937
+ "loss": 1.895,
938
+ "step": 775
939
+ },
940
+ {
941
+ "epoch": 1.26,
942
+ "learning_rate": 2.525363845130185e-05,
943
+ "loss": 1.9826,
944
+ "step": 780
945
+ },
946
+ {
947
+ "epoch": 1.27,
948
+ "learning_rate": 2.5087985151594235e-05,
949
+ "loss": 1.9869,
950
+ "step": 785
951
+ },
952
+ {
953
+ "epoch": 1.28,
954
+ "learning_rate": 2.4921957799879076e-05,
955
+ "loss": 1.9325,
956
+ "step": 790
957
+ },
958
+ {
959
+ "epoch": 1.29,
960
+ "learning_rate": 2.4755568601943615e-05,
961
+ "loss": 1.9479,
962
+ "step": 795
963
+ },
964
+ {
965
+ "epoch": 1.3,
966
+ "learning_rate": 2.4588829790176837e-05,
967
+ "loss": 1.9616,
968
+ "step": 800
969
+ },
970
+ {
971
+ "epoch": 1.3,
972
+ "learning_rate": 2.4421753622670178e-05,
973
+ "loss": 1.9706,
974
+ "step": 805
975
+ },
976
+ {
977
+ "epoch": 1.31,
978
+ "learning_rate": 2.425435238231638e-05,
979
+ "loss": 1.9675,
980
+ "step": 810
981
+ },
982
+ {
983
+ "epoch": 1.32,
984
+ "learning_rate": 2.4086638375906484e-05,
985
+ "loss": 1.9684,
986
+ "step": 815
987
+ },
988
+ {
989
+ "epoch": 1.33,
990
+ "learning_rate": 2.3918623933225043e-05,
991
+ "loss": 1.9388,
992
+ "step": 820
993
+ },
994
+ {
995
+ "epoch": 1.34,
996
+ "learning_rate": 2.375032140614372e-05,
997
+ "loss": 1.9326,
998
+ "step": 825
999
+ },
1000
+ {
1001
+ "epoch": 1.34,
1002
+ "learning_rate": 2.3581743167713187e-05,
1003
+ "loss": 1.9521,
1004
+ "step": 830
1005
+ },
1006
+ {
1007
+ "epoch": 1.35,
1008
+ "learning_rate": 2.3412901611253524e-05,
1009
+ "loss": 1.9704,
1010
+ "step": 835
1011
+ },
1012
+ {
1013
+ "epoch": 1.36,
1014
+ "learning_rate": 2.3243809149443077e-05,
1015
+ "loss": 1.89,
1016
+ "step": 840
1017
+ },
1018
+ {
1019
+ "epoch": 1.37,
1020
+ "learning_rate": 2.3074478213405937e-05,
1021
+ "loss": 1.9438,
1022
+ "step": 845
1023
+ },
1024
+ {
1025
+ "epoch": 1.38,
1026
+ "learning_rate": 2.2904921251798052e-05,
1027
+ "loss": 1.9682,
1028
+ "step": 850
1029
+ },
1030
+ {
1031
+ "epoch": 1.39,
1032
+ "learning_rate": 2.2735150729892013e-05,
1033
+ "loss": 2.008,
1034
+ "step": 855
1035
+ },
1036
+ {
1037
+ "epoch": 1.39,
1038
+ "learning_rate": 2.2565179128660667e-05,
1039
+ "loss": 1.9247,
1040
+ "step": 860
1041
+ },
1042
+ {
1043
+ "epoch": 1.4,
1044
+ "learning_rate": 2.2395018943859558e-05,
1045
+ "loss": 1.9377,
1046
+ "step": 865
1047
+ },
1048
+ {
1049
+ "epoch": 1.41,
1050
+ "learning_rate": 2.222468268510828e-05,
1051
+ "loss": 1.9396,
1052
+ "step": 870
1053
+ },
1054
+ {
1055
+ "epoch": 1.42,
1056
+ "learning_rate": 2.2054182874970808e-05,
1057
+ "loss": 1.9848,
1058
+ "step": 875
1059
+ },
1060
+ {
1061
+ "epoch": 1.43,
1062
+ "learning_rate": 2.188353204803486e-05,
1063
+ "loss": 1.9382,
1064
+ "step": 880
1065
+ },
1066
+ {
1067
+ "epoch": 1.43,
1068
+ "learning_rate": 2.1712742749990444e-05,
1069
+ "loss": 1.9431,
1070
+ "step": 885
1071
+ },
1072
+ {
1073
+ "epoch": 1.44,
1074
+ "learning_rate": 2.154182753670749e-05,
1075
+ "loss": 1.9833,
1076
+ "step": 890
1077
+ },
1078
+ {
1079
+ "epoch": 1.45,
1080
+ "learning_rate": 2.1370798973312813e-05,
1081
+ "loss": 1.9338,
1082
+ "step": 895
1083
+ },
1084
+ {
1085
+ "epoch": 1.46,
1086
+ "learning_rate": 2.1199669633266353e-05,
1087
+ "loss": 1.9543,
1088
+ "step": 900
1089
+ },
1090
+ {
1091
+ "epoch": 1.47,
1092
+ "learning_rate": 2.102845209743682e-05,
1093
+ "loss": 1.9455,
1094
+ "step": 905
1095
+ },
1096
+ {
1097
+ "epoch": 1.47,
1098
+ "learning_rate": 2.085715895317679e-05,
1099
+ "loss": 1.9533,
1100
+ "step": 910
1101
+ },
1102
+ {
1103
+ "epoch": 1.48,
1104
+ "learning_rate": 2.0685802793397317e-05,
1105
+ "loss": 2.0128,
1106
+ "step": 915
1107
+ },
1108
+ {
1109
+ "epoch": 1.49,
1110
+ "learning_rate": 2.051439621564216e-05,
1111
+ "loss": 1.9471,
1112
+ "step": 920
1113
+ },
1114
+ {
1115
+ "epoch": 1.5,
1116
+ "learning_rate": 2.0342951821161648e-05,
1117
+ "loss": 1.9474,
1118
+ "step": 925
1119
+ },
1120
+ {
1121
+ "epoch": 1.51,
1122
+ "learning_rate": 2.017148221398625e-05,
1123
+ "loss": 1.9946,
1124
+ "step": 930
1125
+ },
1126
+ {
1127
+ "epoch": 1.51,
1128
+ "learning_rate": 2e-05,
1129
+ "loss": 1.913,
1130
+ "step": 935
1131
+ },
1132
+ {
1133
+ "epoch": 1.52,
1134
+ "learning_rate": 1.9828517786013752e-05,
1135
+ "loss": 1.981,
1136
+ "step": 940
1137
+ },
1138
+ {
1139
+ "epoch": 1.53,
1140
+ "learning_rate": 1.965704817883836e-05,
1141
+ "loss": 1.9809,
1142
+ "step": 945
1143
+ },
1144
+ {
1145
+ "epoch": 1.54,
1146
+ "learning_rate": 1.948560378435784e-05,
1147
+ "loss": 1.9793,
1148
+ "step": 950
1149
+ },
1150
+ {
1151
+ "epoch": 1.55,
1152
+ "learning_rate": 1.9314197206602693e-05,
1153
+ "loss": 1.9207,
1154
+ "step": 955
1155
+ },
1156
+ {
1157
+ "epoch": 1.56,
1158
+ "learning_rate": 1.914284104682322e-05,
1159
+ "loss": 1.8926,
1160
+ "step": 960
1161
+ },
1162
+ {
1163
+ "epoch": 1.56,
1164
+ "learning_rate": 1.897154790256319e-05,
1165
+ "loss": 2.0005,
1166
+ "step": 965
1167
+ },
1168
+ {
1169
+ "epoch": 1.57,
1170
+ "learning_rate": 1.8800330366733654e-05,
1171
+ "loss": 1.9432,
1172
+ "step": 970
1173
+ },
1174
+ {
1175
+ "epoch": 1.58,
1176
+ "learning_rate": 1.862920102668719e-05,
1177
+ "loss": 1.8667,
1178
+ "step": 975
1179
+ },
1180
+ {
1181
+ "epoch": 1.59,
1182
+ "learning_rate": 1.8458172463292516e-05,
1183
+ "loss": 1.9405,
1184
+ "step": 980
1185
+ },
1186
+ {
1187
+ "epoch": 1.6,
1188
+ "learning_rate": 1.828725725000956e-05,
1189
+ "loss": 1.9617,
1190
+ "step": 985
1191
+ },
1192
+ {
1193
+ "epoch": 1.6,
1194
+ "learning_rate": 1.8116467951965145e-05,
1195
+ "loss": 1.9447,
1196
+ "step": 990
1197
+ },
1198
+ {
1199
+ "epoch": 1.61,
1200
+ "learning_rate": 1.79458171250292e-05,
1201
+ "loss": 1.9093,
1202
+ "step": 995
1203
+ },
1204
+ {
1205
+ "epoch": 1.62,
1206
+ "learning_rate": 1.7775317314891724e-05,
1207
+ "loss": 1.9051,
1208
+ "step": 1000
1209
+ },
1210
+ {
1211
+ "epoch": 1.63,
1212
+ "learning_rate": 1.7604981056140446e-05,
1213
+ "loss": 1.916,
1214
+ "step": 1005
1215
+ },
1216
+ {
1217
+ "epoch": 1.64,
1218
+ "learning_rate": 1.7434820871339336e-05,
1219
+ "loss": 1.8569,
1220
+ "step": 1010
1221
+ },
1222
+ {
1223
+ "epoch": 1.64,
1224
+ "learning_rate": 1.7264849270107994e-05,
1225
+ "loss": 1.9163,
1226
+ "step": 1015
1227
+ },
1228
+ {
1229
+ "epoch": 1.65,
1230
+ "learning_rate": 1.709507874820195e-05,
1231
+ "loss": 1.9342,
1232
+ "step": 1020
1233
+ },
1234
+ {
1235
+ "epoch": 1.66,
1236
+ "learning_rate": 1.6925521786594067e-05,
1237
+ "loss": 1.8947,
1238
+ "step": 1025
1239
+ },
1240
+ {
1241
+ "epoch": 1.67,
1242
+ "learning_rate": 1.675619085055693e-05,
1243
+ "loss": 1.9396,
1244
+ "step": 1030
1245
+ },
1246
+ {
1247
+ "epoch": 1.68,
1248
+ "learning_rate": 1.6587098388746486e-05,
1249
+ "loss": 1.9416,
1250
+ "step": 1035
1251
+ },
1252
+ {
1253
+ "epoch": 1.68,
1254
+ "learning_rate": 1.6418256832286816e-05,
1255
+ "loss": 1.9382,
1256
+ "step": 1040
1257
+ },
1258
+ {
1259
+ "epoch": 1.69,
1260
+ "learning_rate": 1.6249678593856288e-05,
1261
+ "loss": 1.9747,
1262
+ "step": 1045
1263
+ },
1264
+ {
1265
+ "epoch": 1.7,
1266
+ "learning_rate": 1.6081376066774964e-05,
1267
+ "loss": 1.8799,
1268
+ "step": 1050
1269
+ },
1270
+ {
1271
+ "epoch": 1.71,
1272
+ "learning_rate": 1.591336162409352e-05,
1273
+ "loss": 1.8957,
1274
+ "step": 1055
1275
+ },
1276
+ {
1277
+ "epoch": 1.72,
1278
+ "learning_rate": 1.5745647617683627e-05,
1279
+ "loss": 1.8921,
1280
+ "step": 1060
1281
+ },
1282
+ {
1283
+ "epoch": 1.73,
1284
+ "learning_rate": 1.557824637732983e-05,
1285
+ "loss": 1.9406,
1286
+ "step": 1065
1287
+ },
1288
+ {
1289
+ "epoch": 1.73,
1290
+ "learning_rate": 1.5411170209823177e-05,
1291
+ "loss": 1.9282,
1292
+ "step": 1070
1293
+ },
1294
+ {
1295
+ "epoch": 1.74,
1296
+ "learning_rate": 1.5244431398056392e-05,
1297
+ "loss": 1.8621,
1298
+ "step": 1075
1299
+ },
1300
+ {
1301
+ "epoch": 1.75,
1302
+ "learning_rate": 1.5078042200120933e-05,
1303
+ "loss": 1.9375,
1304
+ "step": 1080
1305
+ },
1306
+ {
1307
+ "epoch": 1.76,
1308
+ "learning_rate": 1.4912014848405771e-05,
1309
+ "loss": 1.8779,
1310
+ "step": 1085
1311
+ },
1312
+ {
1313
+ "epoch": 1.77,
1314
+ "learning_rate": 1.4746361548698151e-05,
1315
+ "loss": 1.9353,
1316
+ "step": 1090
1317
+ },
1318
+ {
1319
+ "epoch": 1.77,
1320
+ "learning_rate": 1.4581094479286234e-05,
1321
+ "loss": 1.9255,
1322
+ "step": 1095
1323
+ },
1324
+ {
1325
+ "epoch": 1.78,
1326
+ "learning_rate": 1.4416225790063784e-05,
1327
+ "loss": 1.9163,
1328
+ "step": 1100
1329
+ },
1330
+ {
1331
+ "epoch": 1.79,
1332
+ "learning_rate": 1.4251767601636965e-05,
1333
+ "loss": 1.9314,
1334
+ "step": 1105
1335
+ },
1336
+ {
1337
+ "epoch": 1.8,
1338
+ "learning_rate": 1.4087732004433258e-05,
1339
+ "loss": 1.8751,
1340
+ "step": 1110
1341
+ },
1342
+ {
1343
+ "epoch": 1.81,
1344
+ "learning_rate": 1.3924131057812642e-05,
1345
+ "loss": 1.8934,
1346
+ "step": 1115
1347
+ },
1348
+ {
1349
+ "epoch": 1.81,
1350
+ "learning_rate": 1.376097678918098e-05,
1351
+ "loss": 1.9148,
1352
+ "step": 1120
1353
+ },
1354
+ {
1355
+ "epoch": 1.82,
1356
+ "learning_rate": 1.3598281193105858e-05,
1357
+ "loss": 1.8754,
1358
+ "step": 1125
1359
+ },
1360
+ {
1361
+ "epoch": 1.83,
1362
+ "learning_rate": 1.3436056230434747e-05,
1363
+ "loss": 1.9183,
1364
+ "step": 1130
1365
+ },
1366
+ {
1367
+ "epoch": 1.84,
1368
+ "learning_rate": 1.3274313827415678e-05,
1369
+ "loss": 1.9236,
1370
+ "step": 1135
1371
+ },
1372
+ {
1373
+ "epoch": 1.85,
1374
+ "learning_rate": 1.3113065874820506e-05,
1375
+ "loss": 1.889,
1376
+ "step": 1140
1377
+ },
1378
+ {
1379
+ "epoch": 1.85,
1380
+ "learning_rate": 1.295232422707068e-05,
1381
+ "loss": 1.8898,
1382
+ "step": 1145
1383
+ },
1384
+ {
1385
+ "epoch": 1.86,
1386
+ "learning_rate": 1.2792100701365794e-05,
1387
+ "loss": 1.8991,
1388
+ "step": 1150
1389
+ },
1390
+ {
1391
+ "epoch": 1.87,
1392
+ "learning_rate": 1.2632407076814794e-05,
1393
+ "loss": 1.9559,
1394
+ "step": 1155
1395
+ },
1396
+ {
1397
+ "epoch": 1.88,
1398
+ "learning_rate": 1.2473255093570039e-05,
1399
+ "loss": 1.9048,
1400
+ "step": 1160
1401
+ },
1402
+ {
1403
+ "epoch": 1.89,
1404
+ "learning_rate": 1.2314656451964196e-05,
1405
+ "loss": 1.859,
1406
+ "step": 1165
1407
+ },
1408
+ {
1409
+ "epoch": 1.9,
1410
+ "learning_rate": 1.2156622811650043e-05,
1411
+ "loss": 1.8825,
1412
+ "step": 1170
1413
+ },
1414
+ {
1415
+ "epoch": 1.9,
1416
+ "learning_rate": 1.1999165790743338e-05,
1417
+ "loss": 1.9094,
1418
+ "step": 1175
1419
+ },
1420
+ {
1421
+ "epoch": 1.91,
1422
+ "learning_rate": 1.1842296964968652e-05,
1423
+ "loss": 1.937,
1424
+ "step": 1180
1425
+ },
1426
+ {
1427
+ "epoch": 1.92,
1428
+ "learning_rate": 1.1686027866808394e-05,
1429
+ "loss": 1.8838,
1430
+ "step": 1185
1431
+ },
1432
+ {
1433
+ "epoch": 1.93,
1434
+ "learning_rate": 1.1530369984654936e-05,
1435
+ "loss": 1.9023,
1436
+ "step": 1190
1437
+ },
1438
+ {
1439
+ "epoch": 1.94,
1440
+ "learning_rate": 1.1375334761966074e-05,
1441
+ "loss": 1.9099,
1442
+ "step": 1195
1443
+ },
1444
+ {
1445
+ "epoch": 1.94,
1446
+ "learning_rate": 1.122093359642372e-05,
1447
+ "loss": 1.9058,
1448
+ "step": 1200
1449
+ },
1450
+ {
1451
+ "epoch": 1.95,
1452
+ "learning_rate": 1.1067177839095957e-05,
1453
+ "loss": 1.9359,
1454
+ "step": 1205
1455
+ },
1456
+ {
1457
+ "epoch": 1.96,
1458
+ "learning_rate": 1.0914078793602601e-05,
1459
+ "loss": 1.8897,
1460
+ "step": 1210
1461
+ },
1462
+ {
1463
+ "epoch": 1.97,
1464
+ "learning_rate": 1.0761647715284139e-05,
1465
+ "loss": 1.9341,
1466
+ "step": 1215
1467
+ },
1468
+ {
1469
+ "epoch": 1.98,
1470
+ "learning_rate": 1.0609895810374304e-05,
1471
+ "loss": 1.876,
1472
+ "step": 1220
1473
+ },
1474
+ {
1475
+ "epoch": 1.98,
1476
+ "learning_rate": 1.0458834235176225e-05,
1477
+ "loss": 1.8287,
1478
+ "step": 1225
1479
+ },
1480
+ {
1481
+ "epoch": 1.99,
1482
+ "learning_rate": 1.0308474095242267e-05,
1483
+ "loss": 1.8523,
1484
+ "step": 1230
1485
+ },
1486
+ {
1487
+ "epoch": 2.0,
1488
+ "learning_rate": 1.0128983382202781e-05,
1489
+ "loss": 2.0887,
1490
+ "step": 1235
1491
+ },
1492
+ {
1493
+ "epoch": 2.01,
1494
+ "learning_rate": 9.980205236069665e-06,
1495
+ "loss": 1.4855,
1496
+ "step": 1240
1497
+ },
1498
+ {
1499
+ "epoch": 2.02,
1500
+ "learning_rate": 9.832163712437392e-06,
1501
+ "loss": 1.4915,
1502
+ "step": 1245
1503
+ },
1504
+ {
1505
+ "epoch": 2.03,
1506
+ "learning_rate": 9.684869694834003e-06,
1507
+ "loss": 1.4679,
1508
+ "step": 1250
1509
+ },
1510
+ {
1511
+ "epoch": 2.03,
1512
+ "learning_rate": 9.538334011833363e-06,
1513
+ "loss": 1.4298,
1514
+ "step": 1255
1515
+ },
1516
+ {
1517
+ "epoch": 2.04,
1518
+ "learning_rate": 9.392567436259034e-06,
1519
+ "loss": 1.4018,
1520
+ "step": 1260
1521
+ },
1522
+ {
1523
+ "epoch": 2.05,
1524
+ "learning_rate": 9.247580684392345e-06,
1525
+ "loss": 1.4642,
1526
+ "step": 1265
1527
+ },
1528
+ {
1529
+ "epoch": 2.06,
1530
+ "learning_rate": 9.10338441518453e-06,
1531
+ "loss": 1.4434,
1532
+ "step": 1270
1533
+ },
1534
+ {
1535
+ "epoch": 2.07,
1536
+ "learning_rate": 8.959989229473125e-06,
1537
+ "loss": 1.4574,
1538
+ "step": 1275
1539
+ },
1540
+ {
1541
+ "epoch": 2.07,
1542
+ "learning_rate": 8.817405669202619e-06,
1543
+ "loss": 1.4256,
1544
+ "step": 1280
1545
+ },
1546
+ {
1547
+ "epoch": 2.08,
1548
+ "learning_rate": 8.675644216649478e-06,
1549
+ "loss": 1.4539,
1550
+ "step": 1285
1551
+ },
1552
+ {
1553
+ "epoch": 2.09,
1554
+ "learning_rate": 8.534715293651492e-06,
1555
+ "loss": 1.5016,
1556
+ "step": 1290
1557
+ },
1558
+ {
1559
+ "epoch": 2.1,
1560
+ "learning_rate": 8.39462926084159e-06,
1561
+ "loss": 1.4738,
1562
+ "step": 1295
1563
+ },
1564
+ {
1565
+ "epoch": 2.11,
1566
+ "learning_rate": 8.255396416886194e-06,
1567
+ "loss": 1.4265,
1568
+ "step": 1300
1569
+ },
1570
+ {
1571
+ "epoch": 2.11,
1572
+ "learning_rate": 8.117026997728079e-06,
1573
+ "loss": 1.4235,
1574
+ "step": 1305
1575
+ },
1576
+ {
1577
+ "epoch": 2.12,
1578
+ "learning_rate": 7.979531175833828e-06,
1579
+ "loss": 1.5084,
1580
+ "step": 1310
1581
+ },
1582
+ {
1583
+ "epoch": 2.13,
1584
+ "learning_rate": 7.842919059446046e-06,
1585
+ "loss": 1.4426,
1586
+ "step": 1315
1587
+ },
1588
+ {
1589
+ "epoch": 2.14,
1590
+ "learning_rate": 7.707200691840173e-06,
1591
+ "loss": 1.4797,
1592
+ "step": 1320
1593
+ },
1594
+ {
1595
+ "epoch": 2.15,
1596
+ "learning_rate": 7.572386050586196e-06,
1597
+ "loss": 1.4309,
1598
+ "step": 1325
1599
+ },
1600
+ {
1601
+ "epoch": 2.16,
1602
+ "learning_rate": 7.438485046815078e-06,
1603
+ "loss": 1.4505,
1604
+ "step": 1330
1605
+ },
1606
+ {
1607
+ "epoch": 2.16,
1608
+ "learning_rate": 7.305507524490145e-06,
1609
+ "loss": 1.4734,
1610
+ "step": 1335
1611
+ },
1612
+ {
1613
+ "epoch": 2.17,
1614
+ "learning_rate": 7.1734632596834106e-06,
1615
+ "loss": 1.397,
1616
+ "step": 1340
1617
+ },
1618
+ {
1619
+ "epoch": 2.18,
1620
+ "learning_rate": 7.042361959856825e-06,
1621
+ "loss": 1.4341,
1622
+ "step": 1345
1623
+ },
1624
+ {
1625
+ "epoch": 2.19,
1626
+ "learning_rate": 6.912213263148673e-06,
1627
+ "loss": 1.4599,
1628
+ "step": 1350
1629
+ },
1630
+ {
1631
+ "epoch": 2.2,
1632
+ "learning_rate": 6.783026737664942e-06,
1633
+ "loss": 1.4466,
1634
+ "step": 1355
1635
+ },
1636
+ {
1637
+ "epoch": 2.2,
1638
+ "learning_rate": 6.654811880775973e-06,
1639
+ "loss": 1.4435,
1640
+ "step": 1360
1641
+ },
1642
+ {
1643
+ "epoch": 2.21,
1644
+ "learning_rate": 6.527578118418187e-06,
1645
+ "loss": 1.4597,
1646
+ "step": 1365
1647
+ },
1648
+ {
1649
+ "epoch": 2.22,
1650
+ "learning_rate": 6.401334804401171e-06,
1651
+ "loss": 1.4217,
1652
+ "step": 1370
1653
+ },
1654
+ {
1655
+ "epoch": 2.23,
1656
+ "learning_rate": 6.276091219719984e-06,
1657
+ "loss": 1.4477,
1658
+ "step": 1375
1659
+ },
1660
+ {
1661
+ "epoch": 2.24,
1662
+ "learning_rate": 6.151856571872854e-06,
1663
+ "loss": 1.4716,
1664
+ "step": 1380
1665
+ },
1666
+ {
1667
+ "epoch": 2.24,
1668
+ "learning_rate": 6.028639994184277e-06,
1669
+ "loss": 1.4398,
1670
+ "step": 1385
1671
+ },
1672
+ {
1673
+ "epoch": 2.25,
1674
+ "learning_rate": 5.906450545133564e-06,
1675
+ "loss": 1.4442,
1676
+ "step": 1390
1677
+ },
1678
+ {
1679
+ "epoch": 2.26,
1680
+ "learning_rate": 5.785297207688905e-06,
1681
+ "loss": 1.4506,
1682
+ "step": 1395
1683
+ },
1684
+ {
1685
+ "epoch": 2.27,
1686
+ "learning_rate": 5.665188888646935e-06,
1687
+ "loss": 1.4123,
1688
+ "step": 1400
1689
+ },
1690
+ {
1691
+ "epoch": 2.28,
1692
+ "learning_rate": 5.546134417977984e-06,
1693
+ "loss": 1.456,
1694
+ "step": 1405
1695
+ },
1696
+ {
1697
+ "epoch": 2.28,
1698
+ "learning_rate": 5.428142548176876e-06,
1699
+ "loss": 1.4274,
1700
+ "step": 1410
1701
+ },
1702
+ {
1703
+ "epoch": 2.29,
1704
+ "learning_rate": 5.311221953619514e-06,
1705
+ "loss": 1.4062,
1706
+ "step": 1415
1707
+ },
1708
+ {
1709
+ "epoch": 2.3,
1710
+ "learning_rate": 5.195381229925156e-06,
1711
+ "loss": 1.427,
1712
+ "step": 1420
1713
+ },
1714
+ {
1715
+ "epoch": 2.31,
1716
+ "learning_rate": 5.080628893324475e-06,
1717
+ "loss": 1.4783,
1718
+ "step": 1425
1719
+ },
1720
+ {
1721
+ "epoch": 2.32,
1722
+ "learning_rate": 4.9669733800334955e-06,
1723
+ "loss": 1.4356,
1724
+ "step": 1430
1725
+ },
1726
+ {
1727
+ "epoch": 2.33,
1728
+ "learning_rate": 4.854423045633392e-06,
1729
+ "loss": 1.4809,
1730
+ "step": 1435
1731
+ },
1732
+ {
1733
+ "epoch": 2.33,
1734
+ "learning_rate": 4.742986164456196e-06,
1735
+ "loss": 1.4079,
1736
+ "step": 1440
1737
+ },
1738
+ {
1739
+ "epoch": 2.34,
1740
+ "learning_rate": 4.632670928976501e-06,
1741
+ "loss": 1.4884,
1742
+ "step": 1445
1743
+ },
1744
+ {
1745
+ "epoch": 2.35,
1746
+ "learning_rate": 4.523485449209195e-06,
1747
+ "loss": 1.4499,
1748
+ "step": 1450
1749
+ },
1750
+ {
1751
+ "epoch": 2.36,
1752
+ "learning_rate": 4.415437752113223e-06,
1753
+ "loss": 1.4065,
1754
+ "step": 1455
1755
+ },
1756
+ {
1757
+ "epoch": 2.37,
1758
+ "learning_rate": 4.308535781001457e-06,
1759
+ "loss": 1.4888,
1760
+ "step": 1460
1761
+ },
1762
+ {
1763
+ "epoch": 2.37,
1764
+ "learning_rate": 4.202787394956769e-06,
1765
+ "loss": 1.4707,
1766
+ "step": 1465
1767
+ },
1768
+ {
1769
+ "epoch": 2.38,
1770
+ "learning_rate": 4.0982003682542146e-06,
1771
+ "loss": 1.4426,
1772
+ "step": 1470
1773
+ },
1774
+ {
1775
+ "epoch": 2.39,
1776
+ "learning_rate": 3.994782389789535e-06,
1777
+ "loss": 1.3991,
1778
+ "step": 1475
1779
+ },
1780
+ {
1781
+ "epoch": 2.4,
1782
+ "learning_rate": 3.892541062513853e-06,
1783
+ "loss": 1.4187,
1784
+ "step": 1480
1785
+ },
1786
+ {
1787
+ "epoch": 2.41,
1788
+ "learning_rate": 3.7914839028747507e-06,
1789
+ "loss": 1.4248,
1790
+ "step": 1485
1791
+ },
1792
+ {
1793
+ "epoch": 2.41,
1794
+ "learning_rate": 3.691618340263701e-06,
1795
+ "loss": 1.447,
1796
+ "step": 1490
1797
+ },
1798
+ {
1799
+ "epoch": 2.42,
1800
+ "learning_rate": 3.5929517164698436e-06,
1801
+ "loss": 1.4394,
1802
+ "step": 1495
1803
+ },
1804
+ {
1805
+ "epoch": 2.43,
1806
+ "learning_rate": 3.495491285140282e-06,
1807
+ "loss": 1.4359,
1808
+ "step": 1500
1809
+ },
1810
+ {
1811
+ "epoch": 2.44,
1812
+ "learning_rate": 3.399244211246779e-06,
1813
+ "loss": 1.4752,
1814
+ "step": 1505
1815
+ },
1816
+ {
1817
+ "epoch": 2.45,
1818
+ "learning_rate": 3.304217570559052e-06,
1819
+ "loss": 1.4508,
1820
+ "step": 1510
1821
+ },
1822
+ {
1823
+ "epoch": 2.45,
1824
+ "learning_rate": 3.2104183491245466e-06,
1825
+ "loss": 1.4718,
1826
+ "step": 1515
1827
+ },
1828
+ {
1829
+ "epoch": 2.46,
1830
+ "learning_rate": 3.117853442754879e-06,
1831
+ "loss": 1.4514,
1832
+ "step": 1520
1833
+ },
1834
+ {
1835
+ "epoch": 2.47,
1836
+ "learning_rate": 3.026529656518864e-06,
1837
+ "loss": 1.399,
1838
+ "step": 1525
1839
+ },
1840
+ {
1841
+ "epoch": 2.48,
1842
+ "learning_rate": 2.936453704242215e-06,
1843
+ "loss": 1.4136,
1844
+ "step": 1530
1845
+ },
1846
+ {
1847
+ "epoch": 2.49,
1848
+ "learning_rate": 2.8476322080139862e-06,
1849
+ "loss": 1.4474,
1850
+ "step": 1535
1851
+ },
1852
+ {
1853
+ "epoch": 2.5,
1854
+ "learning_rate": 2.760071697699729e-06,
1855
+ "loss": 1.4542,
1856
+ "step": 1540
1857
+ },
1858
+ {
1859
+ "epoch": 2.5,
1860
+ "learning_rate": 2.673778610461448e-06,
1861
+ "loss": 1.4176,
1862
+ "step": 1545
1863
+ },
1864
+ {
1865
+ "epoch": 2.51,
1866
+ "learning_rate": 2.588759290284337e-06,
1867
+ "loss": 1.4471,
1868
+ "step": 1550
1869
+ },
1870
+ {
1871
+ "epoch": 2.52,
1872
+ "learning_rate": 2.505019987510426e-06,
1873
+ "loss": 1.4217,
1874
+ "step": 1555
1875
+ },
1876
+ {
1877
+ "epoch": 2.53,
1878
+ "learning_rate": 2.4225668583790474e-06,
1879
+ "loss": 1.4194,
1880
+ "step": 1560
1881
+ },
1882
+ {
1883
+ "epoch": 2.54,
1884
+ "learning_rate": 2.3414059645742504e-06,
1885
+ "loss": 1.3959,
1886
+ "step": 1565
1887
+ },
1888
+ {
1889
+ "epoch": 2.54,
1890
+ "learning_rate": 2.261543272779192e-06,
1891
+ "loss": 1.4689,
1892
+ "step": 1570
1893
+ },
1894
+ {
1895
+ "epoch": 2.55,
1896
+ "learning_rate": 2.1829846542374565e-06,
1897
+ "loss": 1.4568,
1898
+ "step": 1575
1899
+ },
1900
+ {
1901
+ "epoch": 2.56,
1902
+ "learning_rate": 2.105735884321436e-06,
1903
+ "loss": 1.451,
1904
+ "step": 1580
1905
+ },
1906
+ {
1907
+ "epoch": 2.57,
1908
+ "learning_rate": 2.029802642107734e-06,
1909
+ "loss": 1.4418,
1910
+ "step": 1585
1911
+ },
1912
+ {
1913
+ "epoch": 2.58,
1914
+ "learning_rate": 1.9551905099596813e-06,
1915
+ "loss": 1.4619,
1916
+ "step": 1590
1917
+ },
1918
+ {
1919
+ "epoch": 2.58,
1920
+ "learning_rate": 1.8819049731169059e-06,
1921
+ "loss": 1.4182,
1922
+ "step": 1595
1923
+ },
1924
+ {
1925
+ "epoch": 2.59,
1926
+ "learning_rate": 1.809951419292104e-06,
1927
+ "loss": 1.4095,
1928
+ "step": 1600
1929
+ },
1930
+ {
1931
+ "epoch": 2.6,
1932
+ "learning_rate": 1.7393351382749424e-06,
1933
+ "loss": 1.4397,
1934
+ "step": 1605
1935
+ },
1936
+ {
1937
+ "epoch": 2.61,
1938
+ "learning_rate": 1.6700613215431549e-06,
1939
+ "loss": 1.4747,
1940
+ "step": 1610
1941
+ },
1942
+ {
1943
+ "epoch": 2.62,
1944
+ "learning_rate": 1.6021350618809184e-06,
1945
+ "loss": 1.4356,
1946
+ "step": 1615
1947
+ },
1948
+ {
1949
+ "epoch": 2.62,
1950
+ "learning_rate": 1.5355613530044089e-06,
1951
+ "loss": 1.4381,
1952
+ "step": 1620
1953
+ },
1954
+ {
1955
+ "epoch": 2.63,
1956
+ "learning_rate": 1.470345089194709e-06,
1957
+ "loss": 1.4444,
1958
+ "step": 1625
1959
+ },
1960
+ {
1961
+ "epoch": 2.64,
1962
+ "learning_rate": 1.4064910649379803e-06,
1963
+ "loss": 1.469,
1964
+ "step": 1630
1965
+ },
1966
+ {
1967
+ "epoch": 2.65,
1968
+ "learning_rate": 1.3440039745729894e-06,
1969
+ "loss": 1.4427,
1970
+ "step": 1635
1971
+ },
1972
+ {
1973
+ "epoch": 2.66,
1974
+ "learning_rate": 1.2828884119460105e-06,
1975
+ "loss": 1.3941,
1976
+ "step": 1640
1977
+ },
1978
+ {
1979
+ "epoch": 2.67,
1980
+ "learning_rate": 1.2231488700730742e-06,
1981
+ "loss": 1.4452,
1982
+ "step": 1645
1983
+ },
1984
+ {
1985
+ "epoch": 2.67,
1986
+ "learning_rate": 1.1647897408096886e-06,
1987
+ "loss": 1.4236,
1988
+ "step": 1650
1989
+ },
1990
+ {
1991
+ "epoch": 2.68,
1992
+ "learning_rate": 1.107815314527929e-06,
1993
+ "loss": 1.4538,
1994
+ "step": 1655
1995
+ },
1996
+ {
1997
+ "epoch": 2.69,
1998
+ "learning_rate": 1.0522297798010594e-06,
1999
+ "loss": 1.4112,
2000
+ "step": 1660
2001
+ },
2002
+ {
2003
+ "epoch": 2.7,
2004
+ "learning_rate": 9.980372230955693e-07,
2005
+ "loss": 1.4808,
2006
+ "step": 1665
2007
+ },
2008
+ {
2009
+ "epoch": 2.71,
2010
+ "learning_rate": 9.452416284707743e-07,
2011
+ "loss": 1.4509,
2012
+ "step": 1670
2013
+ },
2014
+ {
2015
+ "epoch": 2.71,
2016
+ "learning_rate": 8.938468772859132e-07,
2017
+ "loss": 1.4414,
2018
+ "step": 1675
2019
+ },
2020
+ {
2021
+ "epoch": 2.72,
2022
+ "learning_rate": 8.438567479147975e-07,
2023
+ "loss": 1.4203,
2024
+ "step": 1680
2025
+ },
2026
+ {
2027
+ "epoch": 2.73,
2028
+ "learning_rate": 7.952749154680405e-07,
2029
+ "loss": 1.4294,
2030
+ "step": 1685
2031
+ },
2032
+ {
2033
+ "epoch": 2.74,
2034
+ "learning_rate": 7.481049515228811e-07,
2035
+ "loss": 1.4136,
2036
+ "step": 1690
2037
+ },
2038
+ {
2039
+ "epoch": 2.75,
2040
+ "learning_rate": 7.023503238606122e-07,
2041
+ "loss": 1.4316,
2042
+ "step": 1695
2043
+ },
2044
+ {
2045
+ "epoch": 2.75,
2046
+ "learning_rate": 6.580143962116281e-07,
2047
+ "loss": 1.4645,
2048
+ "step": 1700
2049
+ },
2050
+ {
2051
+ "epoch": 2.76,
2052
+ "learning_rate": 6.151004280081574e-07,
2053
+ "loss": 1.4692,
2054
+ "step": 1705
2055
+ },
2056
+ {
2057
+ "epoch": 2.77,
2058
+ "learning_rate": 5.736115741446146e-07,
2059
+ "loss": 1.4408,
2060
+ "step": 1710
2061
+ },
2062
+ {
2063
+ "epoch": 2.78,
2064
+ "learning_rate": 5.335508847456794e-07,
2065
+ "loss": 1.4552,
2066
+ "step": 1715
2067
+ },
2068
+ {
2069
+ "epoch": 2.79,
2070
+ "learning_rate": 4.949213049420576e-07,
2071
+ "loss": 1.4657,
2072
+ "step": 1720
2073
+ },
2074
+ {
2075
+ "epoch": 2.79,
2076
+ "learning_rate": 4.577256746539638e-07,
2077
+ "loss": 1.4189,
2078
+ "step": 1725
2079
+ },
2080
+ {
2081
+ "epoch": 2.8,
2082
+ "learning_rate": 4.2196672838233257e-07,
2083
+ "loss": 1.4573,
2084
+ "step": 1730
2085
+ },
2086
+ {
2087
+ "epoch": 2.81,
2088
+ "learning_rate": 3.876470950078037e-07,
2089
+ "loss": 1.4382,
2090
+ "step": 1735
2091
+ },
2092
+ {
2093
+ "epoch": 2.82,
2094
+ "learning_rate": 3.5476929759743927e-07,
2095
+ "loss": 1.4272,
2096
+ "step": 1740
2097
+ },
2098
+ {
2099
+ "epoch": 2.83,
2100
+ "learning_rate": 3.233357532192494e-07,
2101
+ "loss": 1.4866,
2102
+ "step": 1745
2103
+ },
2104
+ {
2105
+ "epoch": 2.84,
2106
+ "learning_rate": 2.933487727644813e-07,
2107
+ "loss": 1.4132,
2108
+ "step": 1750
2109
+ },
2110
+ {
2111
+ "epoch": 2.84,
2112
+ "learning_rate": 2.648105607777507e-07,
2113
+ "loss": 1.4498,
2114
+ "step": 1755
2115
+ },
2116
+ {
2117
+ "epoch": 2.85,
2118
+ "learning_rate": 2.3772321529494712e-07,
2119
+ "loss": 1.4505,
2120
+ "step": 1760
2121
+ },
2122
+ {
2123
+ "epoch": 2.86,
2124
+ "learning_rate": 2.1208872768901713e-07,
2125
+ "loss": 1.4338,
2126
+ "step": 1765
2127
+ },
2128
+ {
2129
+ "epoch": 2.87,
2130
+ "learning_rate": 1.8790898252354583e-07,
2131
+ "loss": 1.4299,
2132
+ "step": 1770
2133
+ },
2134
+ {
2135
+ "epoch": 2.88,
2136
+ "learning_rate": 1.6518575741421904e-07,
2137
+ "loss": 1.4378,
2138
+ "step": 1775
2139
+ },
2140
+ {
2141
+ "epoch": 2.88,
2142
+ "learning_rate": 1.4392072289814319e-07,
2143
+ "loss": 1.4323,
2144
+ "step": 1780
2145
+ },
2146
+ {
2147
+ "epoch": 2.89,
2148
+ "learning_rate": 1.241154423110169e-07,
2149
+ "loss": 1.4144,
2150
+ "step": 1785
2151
+ },
2152
+ {
2153
+ "epoch": 2.9,
2154
+ "learning_rate": 1.0577137167221863e-07,
2155
+ "loss": 1.4343,
2156
+ "step": 1790
2157
+ },
2158
+ {
2159
+ "epoch": 2.91,
2160
+ "learning_rate": 8.88898595777543e-08,
2161
+ "loss": 1.4625,
2162
+ "step": 1795
2163
+ },
2164
+ {
2165
+ "epoch": 2.92,
2166
+ "learning_rate": 7.347214710111239e-08,
2167
+ "loss": 1.3614,
2168
+ "step": 1800
2169
+ },
2170
+ {
2171
+ "epoch": 2.92,
2172
+ "learning_rate": 5.951936770202782e-08,
2173
+ "loss": 1.4099,
2174
+ "step": 1805
2175
+ },
2176
+ {
2177
+ "epoch": 2.93,
2178
+ "learning_rate": 4.7032547143155417e-08,
2179
+ "loss": 1.4601,
2180
+ "step": 1810
2181
+ },
2182
+ {
2183
+ "epoch": 2.94,
2184
+ "learning_rate": 3.60126034146524e-08,
2185
+ "loss": 1.4231,
2186
+ "step": 1815
2187
+ },
2188
+ {
2189
+ "epoch": 2.95,
2190
+ "learning_rate": 2.6460346666696835e-08,
2191
+ "loss": 1.4549,
2192
+ "step": 1820
2193
+ },
2194
+ {
2195
+ "epoch": 2.96,
2196
+ "learning_rate": 1.8376479149926353e-08,
2197
+ "loss": 1.4122,
2198
+ "step": 1825
2199
+ },
2200
+ {
2201
+ "epoch": 2.96,
2202
+ "learning_rate": 1.176159516380837e-08,
2203
+ "loss": 1.3961,
2204
+ "step": 1830
2205
+ },
2206
+ {
2207
+ "epoch": 2.97,
2208
+ "learning_rate": 6.616181012955025e-09,
2209
+ "loss": 1.4652,
2210
+ "step": 1835
2211
+ },
2212
+ {
2213
+ "epoch": 2.98,
2214
+ "learning_rate": 2.9406149713628874e-09,
2215
+ "loss": 1.425,
2216
+ "step": 1840
2217
+ },
2218
+ {
2219
+ "epoch": 2.99,
2220
+ "learning_rate": 7.351672546129785e-10,
2221
+ "loss": 1.4916,
2222
+ "step": 1845
2223
+ },
2224
+ {
2225
+ "epoch": 3.0,
2226
+ "learning_rate": 0.0,
2227
+ "loss": 1.3883,
2228
+ "step": 1850
2229
+ },
2230
+ {
2231
+ "epoch": 3.0,
2232
+ "step": 1851,
2233
+ "total_flos": 1.8340911973547377e+18,
2234
+ "train_loss": 2.0081952639620475,
2235
+ "train_runtime": 43810.4873,
2236
+ "train_samples_per_second": 5.415,
2237
+ "train_steps_per_second": 0.042
2238
+ }
2239
+ ],
2240
+ "max_steps": 1851,
2241
+ "num_train_epochs": 3,
2242
+ "total_flos": 1.8340911973547377e+18,
2243
+ "trial_name": null,
2244
+ "trial_params": null
2245
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2b8aabfc07bd4f8b579a13e3c5ed3e8dbbef17b9bcbe8603b010a6c6aac8df2d
3
+ size 4207
vocab.json ADDED
The diff for this file is too large to render. See raw diff
 
zero_to_fp32.py ADDED
@@ -0,0 +1,484 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+
3
+ # This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
4
+ # copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
5
+ # the future. Once extracted, the weights don't require DeepSpeed and can be used in any
6
+ # application.
7
+ #
8
+ # example: python zero_to_fp32.py . pytorch_model.bin
9
+
10
+ import argparse
11
+ import torch
12
+ import glob
13
+ import math
14
+ import os
15
+ import re
16
+ from collections import OrderedDict
17
+
18
+ # while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
19
+ # DeepSpeed data structures it has to be available in the current python environment.
20
+ import deepspeed
21
+ from deepspeed.utils import logger
22
+ from deepspeed.checkpoint.constants import (DS_VERSION,
23
+ OPTIMIZER_STATE_DICT,
24
+ PARAM_SHAPES,
25
+ SINGLE_PARTITION_OF_FP32_GROUPS,
26
+ FP32_FLAT_GROUPS,
27
+ ZERO_STAGE,
28
+ PARTITION_COUNT,
29
+ PARAM_SHAPES,
30
+ BUFFER_NAMES)
31
+
32
+ debug = 0
33
+
34
+ # load to cpu
35
+ device = torch.device('cpu')
36
+
37
+
38
+ def atoi(text):
39
+ return int(text) if text.isdigit() else text
40
+
41
+
42
+ def natural_keys(text):
43
+ '''
44
+ alist.sort(key=natural_keys) sorts in human order
45
+ http://nedbatchelder.com/blog/200712/human_sorting.html
46
+ (See Toothy's implementation in the comments)
47
+ '''
48
+ return [atoi(c) for c in re.split(r'(\d+)', text)]
49
+
50
+
51
+ def get_model_state_file(checkpoint_dir, zero_stage):
52
+ if not os.path.isdir(checkpoint_dir):
53
+ raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
54
+
55
+ # there should be only one file
56
+ if zero_stage == 2:
57
+ file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
58
+ elif zero_stage == 3:
59
+ file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
60
+
61
+ if not os.path.exists(file):
62
+ raise FileNotFoundError(f"can't find model states file at '{file}'")
63
+
64
+ return file
65
+
66
+
67
+ def get_optim_files(checkpoint_dir):
68
+ # XXX: need to test that this simple glob rule works for multi-node setup too
69
+ optim_files = sorted(glob.glob(os.path.join(checkpoint_dir,
70
+ "*_optim_states.pt")),
71
+ key=natural_keys)
72
+
73
+ if len(optim_files) == 0:
74
+ raise FileNotFoundError(
75
+ f"can't find '*_optim_states.pt' files in directory '{checkpoint_dir}'")
76
+
77
+ return optim_files
78
+
79
+
80
+ def parse_model_state(file):
81
+ state_dict = torch.load(file, map_location=device)
82
+
83
+ if BUFFER_NAMES not in state_dict:
84
+ raise ValueError(f"{file} is not a model state checkpoint")
85
+ buffer_names = state_dict[BUFFER_NAMES]
86
+ if debug:
87
+ print("Found buffers:", buffer_names)
88
+
89
+ # recover just the buffers while restoring them to fp32 if they were saved in fp16
90
+ buffers = {
91
+ k: v.float()
92
+ for k,
93
+ v in state_dict["module"].items() if k in buffer_names
94
+ }
95
+ param_shapes = state_dict[PARAM_SHAPES]
96
+
97
+ ds_version = state_dict.get(DS_VERSION, None)
98
+
99
+ return buffers, param_shapes, ds_version
100
+
101
+
102
+ def parse_optim_states(files, ds_checkpoint_dir):
103
+
104
+ total_files = len(files)
105
+ state_dicts = []
106
+ for f in files:
107
+ state_dicts.append(torch.load(f, map_location=device))
108
+
109
+ if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
110
+ raise ValueError(f"{files[0]} is not a zero checkpoint")
111
+ zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
112
+ world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
113
+
114
+ # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
115
+ # parameters can be different from data parallelism for non-expert parameters. So we can just
116
+ # use the max of the partition_count to get the dp world_size.
117
+
118
+ if type(world_size) is list:
119
+ world_size = max(world_size)
120
+
121
+ if world_size != total_files:
122
+ raise ValueError(
123
+ f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
124
+ "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
125
+ )
126
+
127
+ # the groups are named differently in each stage
128
+ if zero_stage == 2:
129
+ fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
130
+ elif zero_stage == 3:
131
+ fp32_groups_key = FP32_FLAT_GROUPS
132
+ else:
133
+ raise ValueError(f"unknown zero stage {zero_stage}")
134
+
135
+ if zero_stage == 2:
136
+ fp32_flat_groups = [
137
+ state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key]
138
+ for i in range(len(state_dicts))
139
+ ]
140
+ elif zero_stage == 3:
141
+ # if there is more than one param group, there will be multiple flattened tensors - one
142
+ # flattened tensor per group - for simplicity merge them into a single tensor
143
+ #
144
+ # XXX: could make the script more memory efficient for when there are multiple groups - it
145
+ # will require matching the sub-lists of param_shapes for each param group flattened tensor
146
+
147
+ fp32_flat_groups = [
148
+ torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key],
149
+ 0) for i in range(len(state_dicts))
150
+ ]
151
+
152
+ return zero_stage, world_size, fp32_flat_groups
153
+
154
+
155
+ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
156
+ """
157
+ Returns fp32 state_dict reconstructed from ds checkpoint
158
+
159
+ Args:
160
+ - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
161
+
162
+ """
163
+ print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
164
+
165
+ optim_files = get_optim_files(ds_checkpoint_dir)
166
+ zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
167
+ print(
168
+ f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
169
+
170
+ model_file = get_model_state_file(ds_checkpoint_dir, zero_stage)
171
+ buffers, param_shapes, ds_version = parse_model_state(model_file)
172
+ print(f'Parsing checkpoint created by deepspeed=={ds_version}')
173
+
174
+ if zero_stage == 2:
175
+ return _get_fp32_state_dict_from_zero2_checkpoint(world_size,
176
+ param_shapes,
177
+ fp32_flat_groups,
178
+ buffers)
179
+ elif zero_stage == 3:
180
+ return _get_fp32_state_dict_from_zero3_checkpoint(world_size,
181
+ param_shapes,
182
+ fp32_flat_groups,
183
+ buffers)
184
+
185
+
186
+ def _get_fp32_state_dict_from_zero2_checkpoint(world_size,
187
+ param_shapes,
188
+ fp32_flat_groups,
189
+ buffers):
190
+
191
+ # Reconstruction protocol:
192
+ #
193
+ # XXX: document this
194
+
195
+ if debug:
196
+ for i in range(world_size):
197
+ for j in range(len(fp32_flat_groups[0])):
198
+ print(
199
+ f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
200
+
201
+ # XXX: memory usage doubles here (zero2)
202
+ num_param_groups = len(fp32_flat_groups[0])
203
+ merged_single_partition_of_fp32_groups = []
204
+ for i in range(num_param_groups):
205
+ merged_partitions = [sd[i] for sd in fp32_flat_groups]
206
+ full_single_fp32_vector = torch.cat(merged_partitions, 0)
207
+ merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
208
+ avail_numel = sum([
209
+ full_single_fp32_vector.numel()
210
+ for full_single_fp32_vector in merged_single_partition_of_fp32_groups
211
+ ])
212
+
213
+ if debug:
214
+ wanted_params = sum([len(shapes) for shapes in param_shapes])
215
+ wanted_numel = sum(
216
+ [sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
217
+ # not asserting if there is a mismatch due to possible padding
218
+ print(f"Have {avail_numel} numels to process.")
219
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
220
+
221
+ state_dict = OrderedDict()
222
+
223
+ # buffers
224
+ state_dict.update(buffers)
225
+ if debug:
226
+ print(f"added {len(buffers)} buffers")
227
+
228
+ # params
229
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
230
+ # out-of-core computing solution
231
+ total_numel = 0
232
+ total_params = 0
233
+ for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
234
+ offset = 0
235
+ avail_numel = full_single_fp32_vector.numel()
236
+ for name, shape in shapes.items():
237
+
238
+ unpartitioned_numel = shape.numel()
239
+ total_numel += unpartitioned_numel
240
+ total_params += 1
241
+
242
+ if debug:
243
+ print(
244
+ f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} "
245
+ )
246
+ state_dict[name] = full_single_fp32_vector.narrow(
247
+ 0,
248
+ offset,
249
+ unpartitioned_numel).view(shape)
250
+ offset += unpartitioned_numel
251
+
252
+ # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
253
+ # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
254
+ # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
255
+ # live optimizer object, so we are checking that the numbers are within the right range
256
+ align_to = 2 * world_size
257
+
258
+ def zero2_align(x):
259
+ return align_to * math.ceil(x / align_to)
260
+
261
+ if debug:
262
+ print(f"original offset={offset}, avail_numel={avail_numel}")
263
+
264
+ offset = zero2_align(offset)
265
+ avail_numel = zero2_align(avail_numel)
266
+
267
+ if debug:
268
+ print(f"aligned offset={offset}, avail_numel={avail_numel}")
269
+
270
+ # Sanity check
271
+ if offset != avail_numel:
272
+ raise ValueError(
273
+ f"consumed {offset} numels out of {avail_numel} - something is wrong")
274
+
275
+ print(
276
+ f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
277
+ )
278
+
279
+ return state_dict
280
+
281
+
282
+ def zero3_partitioned_param_info(unpartitioned_numel, world_size):
283
+ remainder = unpartitioned_numel % world_size
284
+ padding_numel = (world_size - remainder) if remainder else 0
285
+ partitioned_numel = math.ceil(unpartitioned_numel / world_size)
286
+ return partitioned_numel, padding_numel
287
+
288
+
289
+ def _get_fp32_state_dict_from_zero3_checkpoint(world_size,
290
+ param_shapes,
291
+ fp32_flat_groups,
292
+ buffers):
293
+
294
+ # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
295
+ # param, re-consolidating each param, while dealing with padding if any
296
+
297
+ avail_numel = fp32_flat_groups[0].numel() * world_size
298
+ # merge list of dicts, preserving order
299
+ param_shapes = {k: v for d in param_shapes for k, v in d.items()}
300
+
301
+ if debug:
302
+ for i in range(world_size):
303
+ print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
304
+
305
+ wanted_params = len(param_shapes)
306
+ wanted_numel = sum(shape.numel() for shape in param_shapes.values())
307
+ # not asserting if there is a mismatch due to possible padding
308
+ print(f"Have {avail_numel} numels to process.")
309
+ print(f"Need {wanted_numel} numels in {wanted_params} params.")
310
+
311
+ state_dict = OrderedDict()
312
+
313
+ # buffers
314
+ state_dict.update(buffers)
315
+ if debug:
316
+ print(f"added {len(buffers)} buffers")
317
+
318
+ # params
319
+ # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
320
+ # out-of-core computing solution
321
+ offset = 0
322
+ total_numel = 0
323
+ total_params = 0
324
+ for name, shape in param_shapes.items():
325
+
326
+ unpartitioned_numel = shape.numel()
327
+ total_numel += unpartitioned_numel
328
+ total_params += 1
329
+
330
+ partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
331
+
332
+ if debug:
333
+ print(
334
+ f"{total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
335
+ )
336
+
337
+ # XXX: memory usage doubles here
338
+ state_dict[name] = torch.cat(
339
+ tuple(fp32_flat_groups[i].narrow(0,
340
+ offset,
341
+ partitioned_numel)
342
+ for i in range(world_size)),
343
+ 0).narrow(0,
344
+ 0,
345
+ unpartitioned_numel).view(shape)
346
+ offset += partitioned_numel
347
+
348
+ offset *= world_size
349
+
350
+ # Sanity check
351
+ if offset != avail_numel:
352
+ raise ValueError(
353
+ f"consumed {offset} numels out of {avail_numel} - something is wrong")
354
+
355
+ print(
356
+ f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements"
357
+ )
358
+
359
+ return state_dict
360
+
361
+
362
+ def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
363
+ """
364
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
365
+ ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
366
+ via a model hub.
367
+
368
+ Args:
369
+ - ``checkpoint_dir``: path to the desired checkpoint folder
370
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
371
+
372
+ Returns:
373
+ - pytorch ``state_dict``
374
+
375
+ Note: this approach may not work if your application doesn't have sufficient free CPU memory and
376
+ you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
377
+ the checkpoint.
378
+
379
+ A typical usage might be ::
380
+
381
+ from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
382
+ # do the training and checkpoint saving
383
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
384
+ model = model.cpu() # move to cpu
385
+ model.load_state_dict(state_dict)
386
+ # submit to model hub or save the model to share with others
387
+
388
+ In this example the ``model`` will no longer be usable in the deepspeed context of the same
389
+ application. i.e. you will need to re-initialize the deepspeed engine, since
390
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
391
+
392
+ If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
393
+
394
+ """
395
+ if tag is None:
396
+ latest_path = os.path.join(checkpoint_dir, 'latest')
397
+ if os.path.isfile(latest_path):
398
+ with open(latest_path, 'r') as fd:
399
+ tag = fd.read().strip()
400
+ else:
401
+ raise ValueError(f"Unable to find 'latest' file at {latest_path}")
402
+
403
+ ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
404
+
405
+ if not os.path.isdir(ds_checkpoint_dir):
406
+ raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
407
+
408
+ return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
409
+
410
+
411
+ def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
412
+ """
413
+ Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
414
+ loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
415
+
416
+ Args:
417
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
418
+ - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
419
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
420
+ """
421
+
422
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
423
+ print(f"Saving fp32 state dict to {output_file}")
424
+ torch.save(state_dict, output_file)
425
+
426
+
427
+ def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
428
+ """
429
+ 1. Put the provided model to cpu
430
+ 2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
431
+ 3. Load it into the provided model
432
+
433
+ Args:
434
+ - ``model``: the model object to update
435
+ - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
436
+ - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
437
+
438
+ Returns:
439
+ - ``model`: modified model
440
+
441
+ Make sure you have plenty of CPU memory available before you call this function. If you don't
442
+ have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
443
+ conveniently placed for you in the checkpoint folder.
444
+
445
+ A typical usage might be ::
446
+
447
+ from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
448
+ model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
449
+ # submit to model hub or save the model to share with others
450
+
451
+ Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
452
+ of the same application. i.e. you will need to re-initialize the deepspeed engine, since
453
+ ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
454
+
455
+ """
456
+ logger.info(f"Extracting fp32 weights")
457
+ state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
458
+
459
+ logger.info(f"Overwriting model with fp32 weights")
460
+ model = model.cpu()
461
+ model.load_state_dict(state_dict, strict=False)
462
+
463
+ return model
464
+
465
+
466
+ if __name__ == "__main__":
467
+
468
+ parser = argparse.ArgumentParser()
469
+ parser.add_argument(
470
+ "checkpoint_dir",
471
+ type=str,
472
+ help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
473
+ parser.add_argument(
474
+ "output_file",
475
+ type=str,
476
+ help=
477
+ "path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
478
+ )
479
+ parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
480
+ args = parser.parse_args()
481
+
482
+ debug = args.debug
483
+
484
+ convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)